Method and apparatus for training an artificial intelligence (AI) model in a wireless network

Abstract

A method and apparatus for training an artificial intelligence AI model in a wireless network are provided. The method includes: sending first setting information to a terminal participating in federated learning, the first setting information being used to set at least one of the following training period, time-frequency resource, and report timing, the same training period, the same time-frequency resource, and the same report timing being set for different terminals participating in federated learning; and receiving a signal obtained by wireless superposition of gradients reported by the terminal, the gradient being completed within the training period and being the gradient of the AI ​​model reported by the terminal at the report timing by using the time-frequency resource. According to the method and apparatus of the present application, problems such as excessively high time-frequency resource utilization and large delay caused by edge nodes participating in federated learning need to use mutually orthogonal uplink time-frequency resources can be eliminated.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of artificial intelligence (AI), and more particularly, to a method for training an AI model in a wireless network.

Background Art

[0002] In wireless communication networks, for example, in mobile communication networks, the services supported by the network are becoming increasingly diverse, so the requirements to be satisfied are also becoming increasingly diverse. For example, the network must have the ability to support ultra-high speed, ultra-low latency, and / or ultra-large capacity connections. Such features make network planning, network configuration, and / or resource scheduling increasingly complex. Such new requirements, scenarios, and features pose unprecedented challenges to network planning, operation, and maintenance, as well as efficient operation. To meet these challenges, artificial intelligence technology can be introduced into wireless communication networks to achieve network intelligence. Federated learning is a popular model training architecture in which distributed data located on edge devices can be invoked to participate in model training without violating privacy. In federated learning, the central node and the edge devices need to exchange the parameters or gradients of the artificial intelligence (AI) model. How to deploy federated learning in a wireless network is a problem worthy of research.

Summary of the Invention

[0003] The present application provides a method and an apparatus for training an AI model in a wireless network so as to solve problems such as excessively high time-frequency resource utilization and large delay caused by the need for edge nodes participating in federated learning to use mutually orthogonal uplink time-frequency resources when federated learning is deployed in the wireless network.

[0004] According to a first aspect, a method for training an artificial intelligence (AI) model in a wireless network is provided. The method may be executed by a second node, may be executed by components (processors, chips, etc.) configured in the second node, or may be executed by a software module. The method includes transmitting first configuration information to terminals participating in federated learning, where the first configuration information is used to set at least one of the following: a training period, time-frequency resources, and reporting timing, and the same training period, the same time-frequency resources, and the same reporting timing are set for different terminals participating in federated learning, and receiving a signal obtained by wireless superposition of gradients reported by terminals participating in federated learning, where the gradients are the gradients reported by the terminals at the reporting timing by using the time-frequency resources of the AI model for which training is completed within the training period.

[0005] It should be noted that during the description of this application, the second node may sometimes be also referred to as an access network device, and the first node may sometimes be also referred to as a central node, etc. In the above design, the same training period, the same time - frequency resource, and the same reporting timing are set for the terminals participating in the federated learning by either the access network device (which may be called the second node) or the central node (which may be called the first node). For example, the number of terminals participating in the federated learning is n. In this application, the access network device or the central node allocates the same time - frequency resource to the n terminals. Compared with the conventional solution where n orthogonal time - frequency resources are allocated to the n terminals participating in the federated learning, the overhead of the time - frequency resource can be reduced. Further, when n orthogonal time - frequency resources are allocated to the n terminals participating in the federated learning, each terminal reports the gradient of the AI model by using each respective time - frequency resource. The access network device can receive n radio frequency signals and process the n radio frequency signals separately to restore the gradients reported by each terminal. As a result, the delay is large. However, in this application, the access network device allocates one time - frequency resource to the n terminals participating in the federated learning, and the n terminals report the gradients with that time - frequency resource. Therefore, based on the property of the superposition of the wireless channels, the n gradients are superimposed during the wireless transmission. For example, the value of n is 3, the gradient reported by terminal 1 is Y1, the gradient reported by terminal 2 is Y2, and the gradient reported by terminal 3 is Y3. In this case, the above three gradients are transmitted with the same time - frequency resource, and the above three gradients are superimposed. It is assumed that the wireless channel satisfies perfect signal superposition (fading, interference, noise, etc. can be ignored). In this case, the superimposed signal is Y = Y1+Y2+Y3.When the channel does not satisfy perfect signal superposition, after receiving the above signal on the above time-frequency resource, the access network device may perform signal processing on the received signal to restore the superimposed signal Y, and the terminal may perform gradient aggregation by using the superimposed signal Y. The process of gradient aggregation may be a process of calculating an arithmetic mean. For example, the superimposed signal Y may be divided by 3, and the result is used as the aggregated gradient. According to the solution method of the present application, the superimposed signal Y can be obtained by processing one radio frequency signal. However, in the existing solution method, three radio frequency signals in different slots need to be sequentially processed to restore the corresponding gradients, and then aggregation is further performed. By using the solution method of the present application, the utilization amount of the time-frequency resource can be reduced to a certain extent, and the delay of gradient aggregation can be reduced.

[0006] It should be understood that the training period set by the second node (or called the access network device) for the terminal may not be the actual training period of the terminal, but the upper limit of the time required by the terminal for each round of training. In other words, the terminal completes the current round of model training within the training period and reports the training completion instruction to the base station. Otherwise, the terminal may end the current round of model training and wait until the next training period arrives. In the present application, the training period is set to ensure that different terminals can report the gradients of model training to the access network device simultaneously. Optionally, the gradient of model training reported by the terminal to the access network device is the gradient in the current round of model training, rather than the gradient in other rounds of model training, for example, the gradient in the previous round of model training. For example, the training period may be a global parameter determined by comprehensively considering the computing power of the terminals participating in federated learning, the complexity of the model, etc.

[0007] In a possible design, the method is to receive a training completion instruction from a terminal, where the training completion instruction is sent by the terminal to a second node when the training of the AI model is completed within the training period, and further includes collecting a statistical value regarding the number of terminals that have completed the training of the AI model within the training period based on the training completion instruction sent by the terminal.

[0008] In a possible design, the method further includes determining an average gradient in the current round of model training based on gradients reported by different terminals participating in federated learning when the number of terminals that have completed the training of the AI model is greater than or equal to a terminal threshold number, or using the average gradient in the previous round of model training as the average gradient in the current round of model training when the number of terminals that have completed the training of the AI model is less than the terminal threshold number, updating the parameters of the AI model based on the average gradient in the current round of model training, and sending the average gradient in the current round of model training to the terminals.

[0009] According to the above design, the first node may determine the terminal threshold number. When aerial computing is introduced into federated learning, the number of terminals participating in federated learning affects the accuracy of calculating the average gradient in the current round of model training. The first node may set the terminal threshold number. When the number of terminals reporting gradients is greater than or equal to the terminal threshold number, the average gradient in the current round of model training is calculated and sent to the terminals, or otherwise, the average gradient in the previous round of model training is sent to the terminals or used as the average gradient in the current round of model training and sent to the terminals to ensure that the accuracy of the calculated average gradient in the current round of model training meets the requirements.

[0010] In a possible design, the method further includes sending to the first node the number of terminals that have completed the training of the AI model within the training period and a signal obtained by wireless superposition of gradients reported by the terminals.

[0011] In a possible design, the first configuration information is further used to configure at least one of the following: dedicated bearer RB resources, modulation scheme, initial AI model, or transmission power.

[0012] In a possible design, the process of determining the transmission power includes measuring the sounding reference signal SRS from the terminal to determine the uplink channel quality of the terminal, and determining the transmission power of the terminal based on the uplink channel quality.

[0013] According to the above design, the second node determines the optimal transmission power by measuring the uplink channel, sets the transmission power for the terminal to transmit the gradient in the current round of model training, improves the accuracy of air computing, and further enhances the accuracy of gradient aggregation.

[0014] In a possible design, the first configuration information is further used to configure at least one of the following: dedicated bearer RB resources, modulation scheme, initial AI model, channel state information CSI interval, or channel inversion parameter.

[0015] In a possible design, the method further includes receiving second configuration information from the first node, and the second configuration information is used to configure at least one of the following: a list of terminals participating in federated learning, an initial AI model, a group temporary identifier, a training period, a terminal threshold number, the size of a transport block, or uplink requirements.

[0016] In a possible design, the method further includes receiving first terminal information from the terminal and transmitting second terminal information to the first node. The first terminal information includes at least one of the following: the communication capability of the terminal, the computing capability of the terminal, or the dataset characteristics of the terminal. The second terminal information includes at least one of the following: the communication capability of the terminal, the computing capability of the terminal, the dataset characteristics of the terminal, or a terminal temporary identifier, and the terminal temporary identifier is assigned to the terminal by the second node.

[0017] In this design, the communication capabilities of the terminal include, for example, the maximum transmission power that the terminal can support, the antenna configuration of the terminal, and the like. The computing capabilities of the terminal include, for example, the performance of the central processing unit (CPU), the performance of the graphics processing unit (GPU), the storage space, and the amount of electricity. The dataset characteristics of the terminal include, for example, the size of the dataset, the distribution of the dataset, whether the dataset is complete, and whether the labels of the dataset are complete. Optionally, the dataset may be further divided into a training set, a validation set, and a test set based on percentages. For example, 60% of the dataset is the training set, 20% of the dataset is the validation set, and 20% of the dataset is the test set. It can be understood that the training set is used to train the AI model, the validation set is used to evaluate the trained AI model, and the test set is used to test the trained AI model. The terminal temporary identifier, that is, the identifier, may be a cell radio network temporary identifier (C-RNTI), other temporary identifiers, and the like.

[0018] In a possible design, the method further includes transmitting a model training end instruction to the terminal when the model training end condition is satisfied, or receiving a model training end instruction from the first node and transmitting the model training end instruction to the terminal.

[0019] According to a second aspect, a method for training an artificial intelligence (AI) model in a wireless network is provided. The method corresponds to the first aspect. For advantageous effects, reference may be made to the description of the first aspect. The method may be executed by a terminal, may be executed by components (such as a processor, a chip, etc.) configured in the terminal, or may be executed by a software module or the like. The method includes receiving first configuration information from a second node, where the first configuration information is used to set at least one of the following: a training period, time-frequency resources, and reporting timing, and the same training period, the same time-frequency resources, and the same reporting timing are set for different terminals participating in federated learning; training the AI model during the training period to obtain the gradient of the AI model in the current round of model training; and reporting the gradient of the AI model in the current round of model training to the second node at the reporting timing by using the time-frequency resources.

[0020] In a possible design, the method further includes, when the training period ends, sending a training completion instruction to the second node if the training of the AI model is completed.

[0021] In a possible design, the method further includes ending the training of the AI model if the training of the AI model is not completed within the training period.

[0022] In a possible design, the method further includes receiving the average gradient in the previous round of model training from the second node and updating the gradient of the AI model in the current round of model training based on the average gradient in the previous round of model training, or updating the gradient and parameters of the AI model in the current round of model training based on the average gradient in the current round of model training.

[0023] In a possible design, the first configuration information is further used to set at least one of the following: dedicated bearer RB resources, modulation scheme, initial AI model, or transmission power.

[0024] In a possible design, the first configuration information is further used to configure at least one of the following: dedicated bearer RB resources, modulation scheme, initial AI model, channel state information CSI interval, or channel inversion parameter.

[0025] In a possible design, when the first configuration information is further used to configure the channel state information CSI interval and the channel inversion parameter, the method further includes: determining the CSI of the terminal's uplink channel based on the measured CSI of the downlink channel if the same frequency resources are configured for the downlink channel and the uplink channel; and determining the transmission power based on the channel inversion parameter if the CSI of the uplink channel meets the requirements of the CSI interval. Reporting the gradient of the AI model in the current round of model training to the second node includes reporting the gradient of the AI model in the current round of model training to the second node based on the determined transmission power.

[0026] In a possible design, the first configuration information further includes a group temporary identifier, which is the group temporary identifier assigned by the first node to the terminal.

[0027] In a possible design, the method further includes receiving a scheduling instruction from the second node, where the scheduling instruction includes a group temporary identifier, and when the group temporary identifier included in the scheduling instruction is the same as the group temporary identifier assigned by the first node to the terminal, performing the training of the AI model in the current round of model training, or skipping the execution of the training of the AI model in the current round of model training if not.

[0028] In a possible design, the method further includes receiving a model training end instruction from the second node and ending the training of the AI model based on the model training end instruction.

[0029] In a possible design, the method further includes transmitting first terminal information to a second node, where the first terminal information includes at least one of the following: the communication capability of the terminal, the computing capability of the terminal, or the dataset characteristics of the terminal.

[0030] According to a third aspect, a method for training an artificial intelligence (AI) model in a wireless network is provided. The method corresponds to the first aspect. For the advantageous effects, refer to the description of the first aspect. The method is executed by a first node and may be executed by a component (processor, chip, or other component) configured in the first node, or may be executed by a software module or the like. The method includes determining second configuration information, where the second configuration information is used to set at least one of the following: a list of terminals participating in federated learning, an initial AI model, a group temporary identifier, a training period, a terminal threshold number, the size of a transport block, or uplink requirements, and transmitting the second configuration information to a second node.

[0031] In a possible design, the method includes receiving second terminal information from the second node, where the second terminal information includes at least one of the following: the communication capability of the terminal, the computing capability of the terminal, the dataset characteristics of the terminal, or a terminal temporary identifier, and the terminal temporary identifier is assigned to the terminal by the second node, and further includes determining a list of terminals participating in federated learning based on the terminal information.

[0032] In a possible design, the method includes receiving, from a second node, a signal obtained by wireless superposition of gradients reported by terminals participating in federated learning and the number of terminals that have completed training of the AI model during a training period; determining an average gradient in the current round of model training based on the gradients of the AI model reported by the terminals when the number of terminals that have completed training of the model during the training period is greater than or equal to a terminal threshold number, or using, otherwise, the average gradient in the previous round of model training as the average gradient in the current round of model training; updating parameters of the AI model based on the average gradient in the current round of model training; and sending the average gradient in the current round of model training to the second node so that the second node can send the average gradient in the current round of model training to the terminals.

[0033] In a possible design, the method further includes sending a scheduling instruction to the second node, where the scheduling instruction includes a group temporary identifier and is used to schedule terminals corresponding to the group temporary identifier to execute training of the AI model in the current round of model training.

[0034] In a possible design, the method further includes sending a model training end instruction to the second node to instruct the terminals to end training of the AI model in the current round of model training when a model training end condition is satisfied.

[0035] According to a fourth aspect, an apparatus is provided. The apparatus includes units or modules that have a one-to-one correspondence with the methods / operations / steps / actions described in the first aspect, the second aspect, or the third aspect. The units or modules may be hardware circuits, or may be software circuits, or may be implemented by hardware circuits in combination with software.

[0036] According to a fifth aspect, a communication device is provided. The device includes a processor and a memory. The memory is configured to store a computer program or instructions, and the processor is coupled to the memory. When the processor executes the computer program or instructions, the device can execute the method in the first aspect, the second aspect, or the third aspect.

[0037] According to a sixth aspect, a device is provided, including a processor and an interface circuit. The processor communicates with other devices through the interface circuit and is configured to execute the method described in any one of the first aspect, the second aspect, or the third aspect.

[0038] According to a seventh aspect, a device is provided, including a processor coupled to a memory. The processor is configured to execute a program stored in the memory to execute the method described in any one of the first aspect, the second aspect, or the third aspect. The memory may be located outside or inside the device. Further, there may be one or more processors.

[0039] According to an eighth aspect, a chip system is provided, including a processor or a circuit configured to execute the method described in any one of the first aspect, the second aspect, or the third aspect.

[0040] According to a ninth aspect, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program or instructions. When the computer program or instructions are executed by a device, the device can execute the method in the first aspect, the second aspect, or the third aspect.

[0041] According to a tenth aspect, a computer program product is provided. The computer program product includes a computer program or instructions. When the computer program or instructions are executed by a device, the device can execute the method in the first aspect, the second aspect, or the third aspect.

[0042] According to the 11th aspect, a system is provided, including a device implementing the method in the 1st aspect and a device executing the method in the 2nd aspect. Optionally, the device may further include a device executing the method in the 3rd aspect.

Brief Description of Drawings

[0043]

Figure 1

Figure 2a

Figure 2b

Figure 2c

Figure 3

Figure 4

Figure 5A

Figure 5B

Figure 6

Figure 7A

Figure 7B

Figure 7C

Figure 8A

Figure 8B

Figure 8C

Figure 8D

Figure 9

Figure 10

Best Mode for Carrying Out the Invention

[0044] FIG. 1 is a diagram of the architecture of a communication system 1000 to which the present application is applied. As shown in FIG. 1, the communication system includes a radio access network 100 and a core network 200. Optionally, the communication system 1000 may further include the Internet 300. The radio access network 100 may include at least one access network device (e.g., 110a and 110b in FIG. 1), and may further include at least one terminal (e.g., 120a to 120j in FIG. 1). The terminal is connected to the access network device in a wireless manner, and the access network device is connected to the core network in a wireless or wired manner. The core network device and the access network device may be different physical devices independent of each other, or the functions of the core network device and the logical functions of the access network device may be incorporated into the same physical device, or a part of the functions of the core network device and a part of the functions of the access network device may be incorporated into one physical device. The terminals may be connected to each other in a wired or wireless manner, and the access network devices may be connected to each other in a wired or wireless manner. FIG. 1 is merely a diagram. The communication system may further include other network devices, for example, it may further include a wireless relay device and a wireless backhaul device, but these are not shown in FIG. 1.

[0045] The access network device may be a base station, an evolved NodeB (eNodeB), a transmission reception point (TRP), a next generation NodeB (gNB) in a 5th generation (5G) mobile communication system, an access network device in an open radio access network (O-RAN), a next generation base station in a 6th generation (6G) mobile communication system, a base station in a future mobile communication system, an access node in a wireless fidelity (Wi-Fi) system, etc., or may be a module or unit, for example, a central unit (CU) that realizes part of the functions of a base station, a distributed unit (DU), a central unit control plane (CU-CP) module, or a central unit user plane (CU-UP) module. The access network device may be a macro base station (e.g., 110a in FIG. 1), or may be a micro base station or an indoor base station (e.g., 110b in FIG. 1), or may be a relay node, a donor node, etc. The specific technologies and specific device forms used by the access network device are not limited in this application.

[0046] In the present application, a device configured to implement the functions of an access network device may be an access network device, or, for example, a device capable of supporting the access network device in implementing functions, such as a chip system, a hardware circuit, a software module, or a combination of a hardware circuit and a software module, where the device may be installed in the access network device or may be adapted to the access network device for use. In the present application, the chip system may include a chip or may include a chip and other discrete components. For the sake of easy description, the following describes the technical solution provided in the present application by using an example in which a device configured to implement the functions of an access network device is an access network device and the access network device is a base station.

[0047] (1) Protocol layer structure

[0048] The communication between the access network device and the terminal follows a specific protocol layer structure. The protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure. For example, the control plane protocol layer structure may include the functions of protocol layers such as a radio resource control (RRC) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, a media access control (MAC) layer, and a physical layer. For example, the user plane protocol layer structure may include the functions of protocol layers such as a PDCP layer, an RLC layer, a MAC layer, and a physical layer. In a possible implementation, a service data adaptation protocol (SDAP) layer may be further included above the PDCP layer.

[0049] Optionally, the protocol layer structure between the access network device and the terminal may further include an artificial intelligence (AI) layer configured to transmit data related to the AI function.

[0050] (2) Central unit (CU) and distributed unit (DU)

[0051] The access device may include a CU and a DU. A plurality of DUs may be controlled by one CU in a centralized manner. For example, the interface between the CU and the DU may be called the F1 interface. The control plane (CP) interface may be F1-C, and the user plane (UP) interface may be F1-U. The specific name of each interface is not limited in this application. The CU and the DU may be defined based on the protocol layer of the wireless network. For example, the functions of the PDCP layer and the protocol layers above the PDCP layer are set by the CU, and the functions of the protocol layers below the PDCP layer (for example, the RLC layer and the MAC layer) are set by the DU. As another example, the functions of the protocol layers above the PDCP layer are set by the CU, and the functions of the PDCP layer and the protocol layers below the PDCP layer are set by the DU. This is not limited.

[0052] The above division of the CU and DU processing functions based on protocol layers is only an example, and alternatively, other divisions may be possible. For example, the CU or DU may be defined to have more protocol layer functions. As another example, the CU or DU may alternatively be defined to have a part of the protocol layer processing functions. In the design, a part of the RLC layer functions and the functions of the protocol layers above the RLC layer are set in the CU, and the remaining functions of the RLC layer and the functions of the protocol layers below the RLC layer are set in the DU. In other designs, the division of the CU or DU functions may alternatively be performed based on service type or other system requirements. For example, the division may be performed based on latency. Functions that need to meet the latency requirements in terms of processing time are set in the DU, and functions that do not need to meet the latency requirements in terms of processing time are set in the CU. In other designs, the CU may alternatively have one or more functions of the core network. For example, the CU may be deployed on the network side to facilitate centralized management. In other designs, the radio unit (RU) of the DU is remotely deployed. Optionally, the RU may have radio frequency functions.

[0053] Optionally, DU and RU can be distinguished at the physical layer (PHY). For example, DU may implement the upper layer functions of the PHY layer, and RU may implement the lower layer functions of the PHY layer. When the PHY layer is used for transmission, the functions of the PHY layer include at least one of the following functions: addition of cyclic redundancy check (CRC) code, channel coding, rate matching, scrambling, modulation, layer mapping, precoding, resource mapping, physical antenna mapping, or radio frequency transmission. When the PHY layer is used for reception, the functions of the PHY layer may include at least one of the following functions: CRC check, channel decoding, derate matching, descrambling, demodulation, layer demapping, channel detection, resource demapping, physical antenna demapping, or radio frequency reception. The upper layer functions of the PHY layer may include a part of the functions of the PHY layer. For example, a part of the functions is closer to the MAC layer. The lower layer functions of the PHY layer may include other parts of the functions of the PHY layer. For example, the said part of the functions is closer to the radio frequency function. For example, the upper layer functions of the PHY layer may include addition of CRC code, channel coding, rate matching, scrambling, modulation, and layer mapping, and the lower layer functions of the PHY layer may include precoding, resource mapping, physical antenna demapping, and radio frequency transmission functions. Alternatively, the upper layer functions of the PHY layer may include addition of CRC code, channel coding, rate matching, scrambling, modulation, layer mapping, and precoding, and the lower layer functions of the PHY layer may include resource mapping, physical antenna mapping, and radio frequency transmission functions. For example, the upper layer functions of the PHY layer may include CRC check, channel decoding, derate matching, decoding, modulation, and layer demapping, and the lower layer functions of the PHY layer may include channel detection, resource demapping, physical antenna demapping, and radio frequency reception functions.Alternatively, the upper layer functions of the PHY layer may include CRC check, channel decoding, derate matching, decoding, demodulation, layer demapping, and channel detection, and the lower layer functions of the PHY layer may include resource demapping, physical antenna demapping, and radio frequency reception functions.

[0054] For example, the functions of the CU may be implemented by one entity, or may be implemented by different entities. For example, the functions of the CU may be further divided. Specifically speaking, the control plane and user plane of the CU are separated and implemented by different entities, and the different entities are a control plane CU entity (i.e., CU-CP entity) and a user plane CU entity (i.e., CU-UP entity). The CU-CP entity and the CU-UP entity may be coupled to the DU so as to jointly realize the functions of the access network device.

[0055] Optionally, any one of the DU, CU, CU-CP, CU-UP, and RU may be a software module, a hardware structure, or a combination of a software module and a hardware structure. This is not limiting. Different entities may exist in different forms, which is not limiting. For example, the DU, CU, CU-CP, and CU-UP are software modules, and the RU is a hardware structure. These modules and the methods executed by these modules are also within the scope of protection of the present disclosure.

[0056] In a possible implementation, the access network device includes a CU-CP, a CU-UP, a DU, and an RU. For example, the present application is executed by the DU, or the DU and the RU, or the CU-CP, the DU, and the RU, or the CU-UP, the DU, and the RU. This is not limiting. The methods executed by the modules are also within the scope of protection of the present application.

[0057] A terminal may also be referred to as a terminal device, user equipment (UE), mobile station, mobile terminal, etc. The terminal is widely applicable to communications in various scenarios including, but not limited to, for example, the following scenarios: device-to-device (D2D), vehicle-to-everything (V2X), machine-type communication (MTC), internet of things (IOT), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, etc. The terminal may be a mobile phone, tablet computer, computer with a wireless transceiver function, wearable device, vehicle, unmanned aerial vehicle, helicopter, airplane, ship, robot, robotic arm, smart home device, etc. The specific technologies and specific device forms used by the terminal are not limited in this application.

[0058] In this application, the device configured to implement the functions of the terminal may be the terminal, or, for example, a device that can support the terminal in implementing the functions, such as a chip system, hardware circuit, software module, or a combination of a hardware circuit and a software module. Here, the device may be installed in the terminal or may be compatible with the terminal for use. For the sake of easy description, the following describes the technical solution provided in this application by using an example where the device configured to implement the functions of the terminal is the terminal.

[0059] The base station and the terminal may be fixed or movable. The base station and / or the terminal may be deployed on the ground, including indoor or outdoor scenarios, and handheld or in-vehicle scenarios, or may be deployed on water, or may be deployed on airplanes, balloons, and artificial satellites in the air. The applicable scenarios of the base station and the terminal are not limited in this application. The base station and the terminal may be deployed in the same scenario or different scenarios. For example, both the base station and the terminal are deployed on the ground. Alternatively, the base station is deployed on the ground and the terminal is deployed on water. Examples are not listed one by one.

[0060] The roles of the base station and the terminal may be relative. For example, the helicopter or unmanned aircraft 120i in FIG. 1 may be configured as a mobile base station. In the case of the terminal 120j accessing the radio access network 100 via 120i, the terminal 120i is a base station, while in the case of the base station 110a, 120i is a terminal. In other words, 110a and 120i communicate with each other based on the radio air interface protocol. 110a and 120i may alternatively communicate with each other based on the interface protocol between base stations. In this case, for 110a, 120i is also a base station. Therefore, the base station and the terminal may be collectively referred to as communication devices. 110a and 110b in FIG. 1 may be called communication devices having the functions of a base station, and 120a to 120j in FIG. 1 may be called communication devices having the functions of a terminal.

[0061] The communication between the base station and the terminal, between the base station and the base station, or between the terminal and the terminal may be carried out using a licensed spectrum, or may be carried out using an unlicensed spectrum, or may be carried out using both the licensed spectrum and the unlicensed spectrum. The communication may be carried out using a spectrum below 6 gigahertz (GHz), or may be carried out using a spectrum above 6 GHz, or may be carried out using both a spectrum below 6 GHz and a spectrum above 6 GHz. The spectrum resources used for wireless communication are not limited in this application.

[0062] In this application, the base station transmits a downlink signal or downlink information to the terminal, the downlink information is carried on the downlink channel, the terminal transmits an uplink signal or uplink information to the base station, and the uplink information is carried on the uplink channel. To communicate with the base station, the terminal may establish a radio connection to a cell controlled by the base station. The cell to which the terminal establishes the radio connection is called the serving cell of the terminal. When communicating with the serving cell, the terminal may be interfered with by signals from nearby cells.

[0063] In this application, an independent network element (e.g., called a central node, an AI network element, or an AI node) may be introduced into the communication system shown in FIG. 1 to implement AI-related operations. The central node may be directly connected to the access network device in the communication system, or may be indirectly connected to the access network device via a third-party network element. The third-party network element may be a core network element such as an authentication management function (AMF) network element or a user plane function (UPF) network element. Alternatively, the AI function, the AI module, or the AI entity may be configured in other network elements in the communication system to implement AI-related operations. For example, the other network element may be an access network device (e.g., gNB), a core network device, or network operation, administration and maintenance (OAM). In this case, the network element that executes the AI-related operation is a network element with a built-in AI function. In this application, OAM is configured to operate, manage, and / or maintain the core network device and / or is configured to operate, manage, and / or maintain the access network device.

[0064] In this application, an AI model is a specific method for implementing AI functions, and the AI model represents the mapping relationship between the input and output of the model. The AI model may be a neural network or other machine learning model. The AI model may sometimes be abbreviated as a model. AI-related operations may include at least one of the following: data collection, model training, model information release, model inference, inference result release, etc.

[0065] A neural network is used as an example. A neural network is a specific implementation form of machine learning technology. According to the universal approximation theorem, a neural network can theoretically approximate any continuous function, whereby the neural network has the function of learning any mapping. In a conventional communication system, a communication module needs to be designed with rich expertise. However, a deep learning communication system based on a neural network can automatically discover implicit pattern structures from a large amount of data sets, establish mapping relationships between data, and obtain performance superior to conventional modeling methods.

[0066] The idea of a neural network is derived from the neuron structure of brain tissue. Each neuron performs a weighted summation operation on the input value of the neuron and outputs the result of the weighted summation through an activation function. Figure 2a is a diagram of the structure of a neuron. The input of the neuron is x = [x0, x1, ···, x n , and it is assumed that the weights corresponding to the inputs are w = [w1, w2, ···, w n , and the bias of the weighted summation is b. The form of the activation function can be diverse. It is assumed that the activation function of one neuron is y = f(z) = max(0, z). In this case, the output of the neuron is

Equation

Number

[0067] A neural network usually includes a multi-layer structure, and each layer contains one or more neurons. As the depth and / or width of the neural network increases, the representation ability of the neural network improves, and it can provide a stronger information extraction ability and abstract modeling ability for complex systems. The depth of the neural network may refer to the number of layers included in the neural network, and the number of neurons included in each layer may be called the width of the layer. Figure 2b is a diagram of the layer relationship of the neural network. In implementation, a neural network includes an input layer and an output layer. After performing neuron processing on the received input, the input layer of the neural network transfers the result to the output layer, and the output layer obtains the output result of the neural network. In other implementations, a neural network includes an input layer, a hidden layer, and an output layer. After performing neuron processing on the received input, the input layer of the neural network transfers the calculation result to the output layer or the adjacent hidden layer, and finally, the output layer obtains the output result of the neural network. A neural network may include one hidden layer or a plurality of sequentially connected hidden layers. This is not limited. In the training process of the neural network, a loss function can be defined. The loss function represents the gap or difference between the output value of the neural network and the ideal target value. The specific form of the loss function is not limited in this application. The training process of the neural network is a process of adjusting neural network parameters such as the gradient of the neural network, the number and width of the layers, the weights of the neurons, the parameters in the activation function of the neurons, and / or the like, so that the value of the loss function is smaller than the threshold or meets the target requirements.

[0068] Figure 2c is a diagram of an AI application framework. The data source is configured to store training data and inference data. The model training host analyzes or trains the training data supplied by the data to obtain an AI model, and deploys the AI model to the model inference host. Optionally, the model training host may further update the AI model deployed on the model inference host. The model inference host may further feedback the relevant information of the deployed model to the model training host so that the model training host can optimize or update the deployed AI model, etc.

[0069] The AI model represents the mapping relationship between the input and output of the model. Obtaining the AI model through learning by the model training host is equivalent to obtaining the mapping relationship between the input and output of the model through learning by the model training host using the training data. The model inference host uses the AI model to perform inference based on the inference data provided by the data source and obtains the inference result. The method may also be described as follows: The model inference host inputs the inference data into the AI model and obtains the output by using the AI model. The output is the inference result. The inference result may indicate the set parameters used (operated) by the subject of the operation and / or the operations performed by the subject of the operation. The inference result may be centrally planned by an actor entity and sent to one or more subjects of the operation (e.g., network entities) for the operation.

[0070] Federated learning (FL) is a popular AI / ML model training framework that can effectively help multiple organizations use data and perform machine learning modeling while satisfying user privacy protection, data security, and government regulation requirements. As a distributed machine learning model, federated learning can effectively solve the problem of data silos, enabling participants to perform joint modeling without sharing data, thereby technically eliminating data silos and realizing AI collaboration. Federated learning includes a central node and edge nodes. The central node (e.g., a server or a base station) can call distributed devices located in edge devices (e.g., smartphones or sensors) to participate in model training without violating privacy.

[0071] Federated learning includes the following three types: horizontal federated learning, vertical federated learning, and federated transfer learning. This application mainly relates to the process of horizontal federated learning. Figure 3 shows the training process of horizontal federated learning. It can be seen that horizontal federated learning includes one central node and multiple edge nodes. The original data is distributed among the edge nodes, the central node does not have the original data, and the edge nodes are not allowed to send the original data to the central node.

[0072] In the training process of federated learning, the central node first sends an initialized AI model (which can be called the initial AI model) to each edge node, and then starts iterative training. Each iterative training process is as follows.

[0073] 1. The edge node trains the initial AI model by using local data and obtains the gradients of the trained AI model.

[0074] 2. Each edge node reports the gradients obtained by training at the edge node to the central node.

[0075] 3. After receiving the gradients reported by the edge nodes, the central node aggregates the gradients and updates the parameters of the AI model based on the aggregated gradients.

[0076] 4. The central node distributes the aggregated gradients to each edge node participating in the training, and the edge nodes update the parameters and gradients of the locally trained AI model based on the aggregated gradients distributed by the central node.

[0077] 5. The central node calculates the loss function of the AI model obtained by updating the parameters. If the loss function meets the conditions, the model training is terminated. If the loss function does not meet the conditions, the above steps 2-4 are repeated.

[0078] In the conventional federated learning solution, mutually orthogonal uplink time-frequency resources need to be allocated to the edge nodes, and the edge nodes transmit the gradients using the mutually orthogonal uplink time-frequency resources. The central node needs to sequentially restore the gradients reported by all the terminals participating in the federated learning training in order to perform gradient aggregation. This causes a large overhead of time-frequency resources and a large delay, and is not applicable to the federated learning scenario with limited bandwidth and / or high delay requirements.

[0079] This application provides a method for training an AI model in a wireless network. In the method, the same time-frequency resources and the same reporting timing can be allocated to the terminals participating in the federated learning. The terminals participating in the federated learning report the gradients of the trained AI model at the same reporting timing by using the same time-frequency resources so as to solve the above problems of large time-frequency resource overhead and large delay caused by allocating a plurality of mutually orthogonal time-frequency resources to the terminals participating in the federated learning.

[0080] As shown in FIG. 4, a procedure for a method of training an AI model in a wireless network is provided. The procedure includes at least the following steps.

[0081] Step 401: The base station transmits first configuration information to the terminals participating in the federated learning. At this time, the first configuration information is used to configure at least one of the following: training period, time-frequency resource, or reporting timing. Correspondingly, the terminal receives the first configuration information from the base station.

[0082] Step 402: The terminal trains the AI model during the training period to obtain the gradient of the AI model in the current round of model training.

[0083] Step 403: The terminal reports the gradient of the AI model in the current round of model training to the base station at the reporting timing by using the time-frequency resource. Correspondingly, the base station receives the signal obtained by wireless superposition of the gradients reported by the terminal.

[0084] In this application, the same training period, the same time-frequency resource, and the same reporting timing are set by the base station or the central node for the terminals participating in the federated learning. For example, the number of terminals participating in the federated learning is n. In this application, the base station or the central node allocates the same time-frequency resource to the n terminals. Compared with the conventional solution in which n orthogonal time-frequency resources are allocated to the n terminals participating in the federated learning, the overhead of the time-frequency resource can be reduced. Further, when n orthogonal time-frequency resources are allocated to the n terminals participating in the federated learning, each terminal reports the gradient of the AI model by using each time-frequency resource. The base station can receive n radio frequency signals and process the n radio frequency signals separately to restore the gradients reported by each terminal. As a result, the delay is large. However, in this application, the base station allocates one time-frequency resource to the n terminals participating in the federated learning, and the n terminals report the gradients with that time-frequency resource. Therefore, based on the superposition property of the radio channel, the n gradients are superimposed during the wireless transmission. For example, the value of n is 3, the gradient reported by terminal 1 is Y1, the gradient reported by terminal 2 is Y2, and the gradient reported by terminal 3 is Y3. In this case, the above three gradients are transmitted with the same time-frequency resource, and the above three gradients are superimposed. It is assumed that the radio channel satisfies perfect signal superposition (fading, interference, noise, etc. can be ignored). In this case, the superimposed signal is Y = Y1 + Y2 + Y3. When the channel does not satisfy perfect signal superposition, after receiving the above signal with the above time-frequency resource, the base station may perform signal processing on the received signal to restore the superimposed signal Y, and the terminal may perform gradient aggregation by using the superimposed signal Y. The process of gradient aggregation may be a process of calculating an arithmetic mean. For example, the superimposed signal Y may be divided by 3, and the result is used as the aggregated gradient. According to the solution of this application, the superimposed signal Y can be obtained by processing one radio frequency signal.However, in existing solutions, the three radio frequency signals in different slots need to be processed sequentially to restore the corresponding gradients, and then aggregation is further performed. By using the solution of the present application, the utilization of time-frequency resources can be reduced to a certain extent, and the delay of gradient aggregation can be reduced.

[0085] The training period set by the base station (or other devices in the access network device) for the terminal may not be the actual training period of the terminal. Usually, it is the upper limit of the time required by the terminal for each round of training. In other words, the terminal completes the current round of model training within the training period and reports the training completion instruction to the base station. Otherwise, the terminal may end the current round of model training and wait until the next training period arrives. In the present application, the base station sets the training period to ensure that different terminals can report the gradients of model training to the base station simultaneously. Optionally, the gradient of model training reported by the terminal to the base station is the gradient in the current round of model training, rather than the gradients in other rounds of model training, for example, the gradients in the previous round of model training. For example, the training period may be determined by comprehensively considering the computing power of the terminals participating in federated learning, the complexity of the model, etc., and then it may be a global parameter set by the base station for each terminal.

[0086] In the present application, the number of terminals participating in federated learning may be n, and each terminal may be regarded as an edge node. The central node may be a base station, or may be an OAM, or may be independently placed as a module in the core network, etc. This is not limited. In the following description, an example where the central node is independent of the base station, in other words, the central node and the base station are two devices, is used for the purpose of explanation.

[0087] As shown in FIGS. 5A and 5B, a procedure for a method of training an AI model in a wireless network is provided. The procedure includes at least the following steps.

[0088] Step 501: Each of the n terminals reports terminal information to the base station. The base station centrally collects the terminal information and reports the collected information to the central node. In this application, the terminal information includes at least one of the following.

[0089] 1. The communication capability of the terminal, including, for example, the maximum transmission power that can be supported by the terminal and the antenna configuration of the terminal.

[0090] 2. The computing capability of the terminal, including, for example, the performance of the central processing unit (CPU), the performance of the graphics processing unit (GPU), the storage space, and the battery level.

[0091] 3. The dataset characteristics of the terminal, which are, for example, the size of the dataset, the distribution of the dataset, whether the dataset is complete, and whether the label of the dataset is complete. Optionally, the dataset may be further divided into a training set, a validation set, and a test set based on a percentage. For example, 60% of the dataset is the training set, 20% of the dataset is the validation set, and 20% of the dataset is the test set. It can be understood that the training set is used to train the AI model, the validation set is used to evaluate the trained AI model, and the test set is used to test the trained AI model.

[0092] Optionally, the base station assigns a terminal temporary identifier to the terminal. The identifier may be a cell radio network temporary identifier (C-RNTI), another temporary identifier, etc. Optionally, the other temporary identifier can be distinguished using a coding method, for example, using the sequence number shown in Table 1 or Table 2. [Table 1]

Table 2

[0093] Optionally, before step 501, the procedure may further include the base station sending an instruction for reporting terminal information to n terminals. The n terminals separately report their respective terminal information in step 501 above based on the instruction.

[0094] Step 502: The central node transmits the second configuration information to the base station. Optionally, the second configuration information is used to configure at least one of the following: a list of terminals participating in the collaborative learning, an initial AI model, a training period, a terminal threshold number, a transport block size, an uplink requirement, etc. The uplink requirement may include a rate, a bit error rate, a delay, etc. during the uplink transmission of the terminal. A transport block is a data block including a MAC protocol data unit (PDU), and the data block is transmitted in a transmission time interval (TTI).

[0095] It should be noted that in this application, when the base station receives the terminal information reported by the terminal, the terminal information may sometimes be referred to as the first terminal information. The base station assigns a temporary identifier to the terminal, and adds the temporary identifier to the terminal information to form the second terminal information. In addition to the terminal temporary identifier, the second terminal information may further include at least one of the following: the communication capability of the terminal, the computing capability of the terminal, or the dataset characteristics of the terminal. The base station reports the second terminal information to the central node.

[0096] Specifically, the central node may determine a list of terminals participating in federated learning based on the second terminal information reported by the base station. For example, the second terminal information includes the communication capabilities, computing capabilities, dataset characteristics, temporary identifiers, etc. of the terminals. For example, the central node comprehensively considers the communication capabilities, computing capabilities, dataset characteristics, etc. of the terminals and generates a list of terminals participating in federated learning. For example, during the comprehensive consideration, priorities may be set for the communication capabilities, computing capabilities, dataset characteristics, etc. of the terminals. The priority of the dataset characteristics of the terminal is higher than the priority of the communication capabilities of the terminal, and the priority of the communication capabilities of the terminal is higher than the priority of the computing capabilities of the terminal. Furthermore, corresponding threshold values are set for the communication capabilities, computing capabilities, and dataset characteristics of the terminals respectively, and terminals that do not meet the threshold conditions are not considered to be placed in the list of terminals participating in federated learning. For example, the central node may use, as terminals participating in federated learning, terminals whose communication capabilities are above the communication capability threshold, computing capabilities are above the computing capability threshold, and dataset characteristics meet the dataset characteristic requirements.

[0097] In this application, the central node may set a training period for the terminals participating in federated learning. The training period needs to be set to ensure that all terminals participating in federated learning can complete local model training within the training period, considering the computational complexity of the AI model to be trained and the computing capabilities of the terminals. However, the training period should not be set too long so as not to affect the overall efficiency of the training of the AI model.

[0098] In this application, the central node can determine the terminal threshold number. When air computing is introduced into federated learning, the number of terminals participating in federated learning affects the accuracy of calculating the average gradient in the current round of model training. Therefore, the central node may set the terminal threshold number. When the number of terminals reporting gradients is greater than or equal to the terminal threshold number, the average gradient in the current round of model training is calculated and sent to the terminals; otherwise, the average gradient in the previous round of model training is sent to the terminals, or the average gradient in the previous round of model training is used as the average gradient in the current round of model training and sent to the terminals.

[0099] Step 503: The base station transmits the first configuration information to the terminal.

[0100] In this application, the first configuration information is used to configure at least one of the following: training period, reporting timing, time-frequency resource, dedicated bearer radio bearer (RB) resource, modulation scheme, initial AI model, etc.

[0101] For example, the base station can determine the time-frequency resource, dedicated bearer RB resource, modulation scheme, etc. based on the uplink requirements. In this application, the same time-frequency resource, the same reporting timing, and the same training period are allocated by the base station to n terminals participating in federated learning.

[0102] For example, the base station needs to allocate dedicated bearer RB resources to the terminal. The dedicated bearer RB resources may be signaling radio bearer (SRB) resources, data radio bearer (DRB) resources, etc., and are used to transmit the gradients of the AI model. The dedicated bearer RB resources may only be used for gradient transmission and cannot be used for other data transmission. The modulation method set by the base station for the terminal may be phase shift keying (PSK), quadrature amplitude modulation (QAM), other modulation methods, etc. The specific order of PSK or QAM used for modulation can be determined based on uplink requirements, uplink channel quality, the communication capabilities of the base station, the communication capabilities of the terminal, etc.

[0103] Step 504: The terminal trains the AI model during the training period.

[0104] In this application, when receiving the first setting information, the terminal may train the AI model. In the first round of the training process, specifically, the terminal trains the initial AI model, and it can be understood that the initial AI model is configured for the terminal by the central node. In the subsequent training process, specifically, the terminal trains the AI model obtained from the previous training. The training period of this application can be represented by T. For the terminal participating in federated learning, if the training of the AI model is completed within the training period T, a training completion instruction is reported to the base station. If the training of the AI model is not completed within the training period T, the model training is terminated.

[0105] Step 505: Based on the training completion instruction reported by the terminal, the base station collects statistical values regarding the number of terminals that have completed the training of the AI model within the training period T, and measures the uplink channel quality of the terminals that have completed the current round of model training. Optionally, for terminals that do not complete model training in the current round of model training, the base station no longer measures the uplink channel quality of the corresponding terminals. Correspondingly, terminals that do not complete model training no longer report the gradients of the AI model in the current round of model training to the base station.

[0106] For example, the base station may set a counter. When each round of training starts, the base station counts the terminals that report the training completion instruction and resets the counter to 0 when each round of training ends. When the training period T ends, if the count of the counter is greater than or equal to the terminal threshold number, the base station measures the uplink channel quality of the terminals, or otherwise, the base station triggers the terminals to execute the next training round.

[0107] Optionally, for one terminal, the process by which the base station measures the uplink channel quality of the terminal includes the base station receiving a sounding reference signal (SRS) from the terminal. The base station measures the SRS to determine the uplink channel quality of the terminal.

[0108] Step 506: Based on the uplink channel quality of the terminal, the base station determines the transmission power of the corresponding terminal, and the base station transmits the third configuration information to the terminal. At this time, the third configuration information is used to configure the transmission power of the terminal, the reporting timing, etc. In this application, the third configuration information and the first configuration information may sometimes be collectively referred to as one piece of configuration information.

[0109] For example, the base station can determine the transmission power of the terminal by comprehensively considering conditions such as the error requirements of air computing, the uplink channel quality of the terminal, the maximum transmission power supported by the terminal, and the total power. Since air computing has requirements regarding the synchronization of nodes, excessive synchronization will affect the accuracy of air computing. Therefore, the base station can set the same reporting timing for the n terminals participating in the federated learning. Furthermore, since the channel is constantly changing, there is a validity period for channel quality measurement. Therefore, the terminal needs to simultaneously report the gradients of the AI model in the current round of model training at the reporting timing, and the reporting timing should be within the validity period of the channel quality measurement.

[0110] Optionally, considering the time overhead for the base station to determine the transmission power of the terminal, the validity period of the channel quality measurement result may be extended, thereby causing inconsistencies between power allocation and uplink channel quality, etc. In the design, the base station can predict the uplink channel quality at the current reporting time based on the past uplink channel quality of the terminal, and optimize the optimal transmission power in advance. The specific prediction method is shown in FIG. 6. The base station collects the past uplink channel quality of each terminal from time point t1 to time point t2, starts predicting the uplink channel quality of each terminal at time point t3 based on the past uplink channel quality of the terminal at time point t2, and distributes the optimal power distribution solution to the terminal at time point t3. After receiving the power distribution solution, the terminal immediately reports the gradient in the current round of model training. worn-out parts quality, and starts predicting the uplink channel quality of each terminal at time point t3 based on the past uplink channel quality of the terminal at time point t2, and distributes the optimal power distribution solution to the terminal at time point t3. After receiving the power distribution solution, the terminal immediately reports the gradient in the current round of model training.

[0111] Step 507: The terminal reports the gradient in the current round of model training, the base station calculates the average gradient in the current round of model training, the base station updates the parameters of the AI model based on the average gradient in the current round of model training, and distributes the average gradient in the current round of model training to the n terminals participating in the federated learning.

[0112] For example, if n terminals participating in federated learning complete the training that is also due within the training period T at the reporting timing, The terminal may report the gradient in the current round of model training to the base station. When receiving the gradients in the current round of model training reported by the n terminals, the base station may calculate the average gradient in the current round of model training based on the gradients reported by the terminals. For example, since the n terminals report the gradients in the current round of model training using the same time-frequency resource, the signal received by the base station is the signal obtained by wireless superposition of the n gradients. For example, if the gradient obtained by superposition of the n gradients is Y, the base station may determine that the average gradient in the current round of model training is Y / n. Optionally, in this application, a terminal that does not complete model training within the training period T no longer reports the gradient in the current round of model training to the base station.

[0113] In this application, in order to reduce the reporting time error of the terminal as much as possible, each terminal should start reporting the gradient in the current round of model training as soon as the reporting timing arrives. The base station needs to process the received signal obtained by air computing to restore the average gradient in the current round of model training. For example, when the count of a counter at the base station is greater than or equal to the terminal threshold number, the base station distributes the locally calculated average gradient in the current round of model training to each terminal. Otherwise, the average gradient in the previous round of model training is used as the average gradient in the current round of model training and distributed to each terminal.

[0114] Optionally, when the terminals participating in federated learning are changed, the base station may send the updated model parameters to the terminals.

[0115] It can be understood that steps 504 to 507 are a cyclic process. After the average gradient in the current round of model training is distributed in step 507, the terminal updates the parameters and gradients of the AI model based on the average gradient in the current round of model training, and may report a training completion instruction to the base station if the model training is completed within the training period T. It should be noted that in machine learning, in order to minimize the loss function, the parameters of the AI model need to descend along the negative direction of the gradient, that is, the gradient descends. In this application, the parameters of the AI model may first be updated in the current round of model training based on the average gradient in the previous round of model training, and then the gradient of the AI model is updated based on the updated parameters, that is, the gradient is updated.

[0116] Step 508: The base station determines the model training end condition and sends a model training end instruction to the terminal.

[0117] For example, the base station may determine the model training end condition and send a model training end instruction to each terminal. The model training end condition may be at least one of the following: the model parameters converge, the maximum number of model training rounds is reached, the maximum time of model training is reached, etc.

[0118] In this application, air computing is introduced to reduce communication delay, overhead of time-frequency resources, and signaling overhead in the collaborative learning process. The set training period and the set reporting timing can reduce the time synchronization error for the terminal to report the gradient. Before the gradient is reported, the base station first measures the uplink channel quality of all participating terminals, optimizes the transmission power of each terminal to improve the performance of air computing, and further improves the training effect of collaborative learning.

[0119] As shown in FIGS. 7A to 7C, the present application provides a procedure for a method of training an AI model in a wireless network. The main difference between this procedure and the procedures shown in FIGS. 5A and 5B is that in this procedure, the terminal determines, by itself, the transmission power for reporting the gradient in the current round of model training, and the transmission power is no longer set by the base station. The procedure includes at least the following steps.

[0120] Step 701: n terminals report terminal information to the base station, and the base station centrally reports the terminal information to the central node.

[0121] Specifically, upon receiving the terminal information of the terminal, base station may assign a temporary identifier to the terminal, add the temporary identifier to the terminal information, and report the terminal information to the central node. See the descriptions of FIGS. 5A and 5B.

[0122] Step 702: The central node transmits second configuration information to the base station. At this time, the second configuration information is used to set at least one of the following: a list of terminals participating in federated learning, an initial AI model, a training period, a terminal threshold number, a channel state information (CSI) interval, a channel inversion parameter, the size of a transport block, or uplink requirements. See the descriptions of FIGS. 5A and 5B for the list of terminals participating in federated learning, the initial AI model, the training period, the terminal threshold number, the size of the transport block, the uplink requirements, etc. This procedure focuses on the CSI interval and the channel inversion parameter.

[0123] In the present application, CSI is channel state information and includes a signal-to-noise ratio, a Doppler frequency shift, a multipath delay spread, etc. The CSI interval includes a signal-to-noise ratio interval, a maximum Doppler frequency shift interval, and a maximum delay spread interval. The signal-to-noise ratio interval is [γ min , γ maxand γ min and γ max are respectively the lower limit and the upper limit of the signal-to-noise ratio. The maximum Doppler frequency shift interval is [f min , f max , where f min and f max are respectively the lower limit and the upper limit of the maximum Doppler frequency shift. The maximum delay spread interval is [τ min , τ max , where τ min and τ max are respectively the lower limit and the upper limit of the maximum delay spread.

[0124] In this application, the gradient in the current round of model training can be reported to the base station only when the downlink CSI of the terminal satisfies the CSI interval. Otherwise, the gradient is not reported. For example, the signal-to-noise ratio, the maximum Doppler frequency shift, and the maximum delay spread measured by the terminal are γ1, f1, and τ1 respectively. When γ min ≤γ1≤γ max , f min ≤f1≤f max , and τ min ≤τ1≤τ max , the terminal can report the gradient in the current round of model training to the base station at the reporting timing. Otherwise, the terminal does not report the gradient in the current round of model training.

[0125] In this application, the channel inversion parameter α is a parameter used for power control. The maximum power and the channel gain of the k-th terminal are P k and h k respectively, and it is assumed that P1|h1| 2 ≤···≤P k |h k | 2 ≤···≤P K |h K | 2 . In this case, α = P1|h1| 2 , and the transmission power used by the k-th terminal to report the gradient is p k =α / |h k |2 should be

[0126] Step 703: The base station determines a time-frequency resource, a dedicated bearer RB resource, a scheduling method, etc. based on the uplink requirements. The base station transmits first configuration information to the terminal, and the first configuration information is used to configure at least one of the following: a training period, a reporting timing, a time-frequency resource, an initial AI model, a dedicated bearer RB resource, a modulation method, etc.

[0127] In this application, the dedicated bearer RB resources include dedicated SRB resources, dedicated DRB resources, and / or the like. The base station may determine the time-frequency resources, SRB / DRB resources, modulation scheme, etc. based on the uplink requirements. To meet the requirements of air computing, all terminals need to use the same time-frequency resources. Optionally, the base station may allocate the same time-frequency resources to n terminals and improve the performance of air computing by using time diversity or frequency diversity. It should be noted that when the base station sets the same time-frequency resources for n terminals, the n terminals report the gradients in the current round of model training to the base station at a specific reporting timing by using the time-frequency resources. For example, when the base station sets three time-frequency resources for n terminals, the n terminals report the gradients in the current round of model training at the reporting timing by simultaneously using the first time-frequency resource, the second time-frequency resource, and the third time-frequency resource among the three time-frequency resources. That is, when the n terminals report the gradients in the current round of model training at the reporting timing, the same time-frequency resources are specifically used. To avoid interference from other data, the base station may allocate independent SRB / DRB resources to the terminals for transmitting the gradients in the current round of model training. The base station may set a modulation scheme such as PSK or QAM for the terminals. The specific order of PSK or QAM used for modulation may be determined based on at least one of the following: uplink requirements, uplink channel quality, communication capabilities of the base station, communication capabilities of the terminals, etc.

[0128] Step 704: The terminal locally trains the model.

[0129] In this application, when the terminal receives the first setting information, it may start to execute model training. If the terminal completes the model training within the training period T, the terminal reports a training completion instruction to the base station. If the terminal does not complete the model training within the training period T, the terminal terminates the model training.

[0130] Step 705: The terminal measures the CSI of the downlink channel and determines the transmission power of the terminal.

[0131] In this application, the same frequency resource can be set for the uplink channel and the downlink channel of the terminal. Based on channel reciprocity, the terminal can measure the CSI of the downlink channel to obtain the CSI of the uplink channel. The terminal can determine whether the obtained CSI of the uplink channel is within the CSI interval. If the obtained CSI of the uplink channel is within the CSI interval, the transmission power is determined based on the channel inversion parameter. From the above description, the channel inversion parameter is α = P1|h1| 2 and the transmission power used by the k-th terminal to report the gradient is p k = α / |h k | 2 It can be seen that it should be so. In this application, for the k-th terminal, the parameter h k can be obtained using CSI, and the transmission power p k of the k-th terminal can be determined with reference to the channel inversion parameter α, where k is a positive integer from 1 to n.

[0132] Step 706: The terminal reports the gradient in the current round of model training, the base station calculates the average gradient in the current round of model training, updates the parameters of the AI model based on the average gradient in the current round of model training, and sends the average gradient in the current round of model training to each terminal participating in the federated learning. For the specific process of Step 706, please refer to Step 507 above.

[0133] Step 707: The base station determines the model end condition and sends a model training end instruction to each terminal.

[0134] It should be noted that in the design, the terminal can use the set training period and the set reporting timing to ensure that the terminal reports the gradients in local training simultaneously. For example, after receiving the average gradient in the round before model training delivered by the base station, the terminal starts to execute the current round of model training and reports the local training gradient in the current round of model training T seconds after receiving the average gradient in the round before model training. T seconds is the training period, and T seconds after receiving the average gradient in the round before model training is the reporting timing.

[0135] In this application, air computing is introduced to reduce communication delay and bandwidth overhead in the federated learning process, and the set training period and the set reporting timing can reduce the time error of the terminal reporting gradients. The base station transmits parameters such as the training period, power adjustment solution, and reporting timing to each terminal participating in the federated learning in a preset manner, and the terminal trains the model periodically and reports the gradients. Before reporting the gradients, the terminal optimizes the transmission power of the terminal through measuring the downlink channel quality by using channel reciprocity, thereby improving the performance of air computing and further improving the training effect of federated learning. Furthermore, since the terminal actively reports the training gradients when the training period ends, the signaling overhead of the base station for scheduling the terminal can be reduced.

[0136] As shown from FIG. 8A to FIG. 8D, a flowchart of a method for training an AI model in a wireless network is provided. The main difference between the flowchart and the procedures shown in FIG. 5A and FIG. 5B is that the terminals participating in the federated learning are grouped, and in that case, the central node schedules the terminals within a specific group to mainly report the gradients in the current round of model training. The flowchart includes at least the following steps.

[0137] Step 801: The terminal reports the terminal information to the base station, and the base station centrally reports the terminal information to the central node.

[0138] For specific descriptions, please refer to the descriptions of FIGS. 5A and 5B.

[0139] Step 802: The central node transmits the second configuration information to the base station.

[0140] In this application, the central node may determine a list of terminals participating in the federated learning, an initial AI model, a training period, a terminal threshold number, a transport block size, an uplink requirement, a group temporary identifier, etc. The second configuration information is used to configure at least one of the following: a list of terminals participating in the federated learning, an initial AI model, a training period, a terminal threshold number, a transport block size, an uplink requirement, a group temporary identifier, etc.

[0141] In this application, the central node comprehensively considers the communication capabilities of the terminals, the computing capabilities of the terminals, dataset characteristics, etc., determines a list of terminals participating in federated learning within the service range of each base station, groups the terminals participating in federated learning and within the service range of the base station into one group, and can assign a temporary identifier to the group. The identifier can be called a group temporary identifier. For example, during the comprehensive consideration, the central node may set priorities for the communication capabilities of the terminals, computing capabilities, dataset characteristics, etc. The priority of the dataset characteristics is higher than the priority of the communication capabilities of the terminals, and the priority of the communication capabilities of the terminals is higher than the priority of the computing capabilities of the terminals. Furthermore, corresponding thresholds are set for the communication capabilities, computing capabilities, and dataset characteristics of the terminals respectively. Terminals that do not meet the threshold conditions are considered not to be placed on the list of terminals participating in federated learning. In this application, the terminals participating in federated learning and within the service range of the base station can be grouped into one group. For example, in the procedures from Figure 8A to Figure 8D, an example is used where m terminals within the coverage area of base station 1 are grouped into one group, and n terminals within the coverage area of base station N are grouped into another group. Both m and n are positive integers, and the values of m and n may be the same or different. For each group, the terminal can assign a temporary identifier called a group temporary identifier to the group. In a possible implementation, two groups are used as an example. Refer to Table 3 below for the list of terminals participating in federated learning in each group.

Table 3

[0142] Step 803: The base station sends the first configuration information to the terminal. The first configuration information is used to configure at least one of the following: the initial AI model, the group temporary identifier, the training period, the time-frequency resource, the dedicated bearer RB resource, the modulation scheme, etc.

[0143] In this application, the base station may determine the time-frequency resource, dedicated bearer RB resource, modulation method, etc. based on the uplink requirements in the first configuration information. Alternatively, the time-frequency resource may be set by the central node and distributed to each terminal through the base station. The central node may allocate the same time-frequency resource to all terminals within the same group.

[0144] Step 804: The central node sends a scheduling instruction to the base station, and the base station forwards the scheduling instruction to the terminal. The scheduling instruction includes a group temporary identifier and is used to schedule the terminal corresponding to the group temporary identifier to perform the training of the AI model in the current round of model training.

[0145] In this application, an example is used in which n terminals within the coverage area of one base station are grouped into one group. When the central node specifically schedules a specific group to perform model training and report the gradient in the current round of model training, the central node sends a scheduling instruction to the base station corresponding to the group. Upon receiving the scheduling instruction, the base station may broadcast the scheduling instruction within the coverage area of the base station. Upon receiving the scheduling instruction, the terminal may compare the group temporary identifier carried in the scheduling instruction with the group temporary identifier assigned to the terminal by the central node. If the two are the same, the terminal performs the training of the AI model in the current round of model training; otherwise, the terminal does not perform the training of the AI model in the current round of model training.

[0146] Similar to the procedure shown in FIGS. 5A and 5B, if the terminal has completed the training of the AI model when the training period ends, the terminal reports a training completion instruction to the base station. If the terminal has not completed the training of the AI model when the training period ends, the terminal also ends the training of the AI model. The base station counts the number of terminals that have completed model training during the training period based on the training completion instruction reported by the terminal, and reports the number of terminals to the central node.

[0147] Step 805: The base station transmits third configuration information to the terminal, and the third configuration information is used to set the transmission power and reporting timing of each terminal. Optionally, the base station sets the same reporting timing for the terminals. Alternatively, the reporting timing may be set by the central node and transferred to each terminal by the base station. In other words, the second configuration information may further be used to set the reporting timing. In this application, the third configuration information and the first configuration information may sometimes be referred to as one configuration information.

[0148] For example, the process of determining the transmission power by the base station includes the base station measuring the SRS from the terminal to determine the uplink channel quality of the terminal. The terminal determines its transmission power based on the uplink channel quality. For details, refer to the descriptions of FIGS. 5A and 5B. The details are not described again here.

[0149] Step 806: The terminal reports the gradient in the current round of model training to the central node. The central node calculates the average gradient in the current round of model training, updates the parameters of the AI model based on the average gradient in the current round of model training, and distributes the average gradient in the current round of model training to the terminals within the current scheduling group through the base station.

[0150] Specifically, the central node can compare the value relationship between the number of terminals that have completed the training of the AI model during the training period and the terminal threshold number. When the number of terminals that have completed the training of the AI model during the training period is greater than the terminal threshold number, the central node calculates the average gradient in the current round of model training based on the gradient in the current round of model training reported by the terminals, or otherwise, the central node uses the average gradient in the previous round of model training as the average gradient in the current round of model training. The central node updates the parameters of the AI model based on the average gradient in the current round of model training and sends the average gradient in the current round of model training to the terminals within the above scheduling group through the base station.

[0151] Step 807: The central node determines the model training end condition and sends a model training end instruction to the corresponding terminal.

[0152] Regarding the model training end condition, please refer to the descriptions of the procedures in FIGS. 5A and 5B or FIGS. 7A to 7C. The difference is that in the descriptions of the procedures in FIGS. 8A to 8D, the central node determines whether the model training end condition is satisfied, while in the procedures shown in FIGS. 5A and 5B or FIGS. 7A to 7C, the base station determines whether the model training end condition is satisfied. It can be understood that the central node may first send a model training end instruction to the base station, and the base station transfers the model training end instruction to the terminal.

[0153] In the procedure shown in FIGS. 8A to 8D, the central node may schedule the terminals within all groups to participate in the federated learning training, or may schedule only the terminals within some groups to participate in the federated learning training. In the above description, when receiving the scheduling instruction, the terminal may execute the federated learning training, or when not, the terminal does not execute the federated learning training. Optionally, the terminal may further set the monitoring parameters of the terminal, and the parameters include the monitoring timing and the monitoring period. When the monitoring timing arrives, the scheduling instruction is monitored based on the set monitoring period. When the monitoring period is reached, the terminal stays in the sleep mode to save the power of the terminal.

[0154] In the above solution, the terminals are grouped, and not all terminals need to be scheduled to execute the federated learning training. Compared with the solution where all terminals need to participate in the federated learning, the power consumption of the terminals can be reduced.

[0155] To implement the functions in the above method, it can be understood that the base station, the terminal, and the central node include corresponding hardware structures and / or software modules for executing the functions. Those skilled in the art should easily notice that the present application can be implemented by hardware or a combination of hardware and computer software with reference to the units and method steps in the examples described in the present application. Whether the function is executed by hardware driven by hardware or computer software depends on the specific application scenario and design constraints of the technical solution.

[0156] Figures 9 and 10 are diagrams of the structures of possible communication devices according to the present application. These communication devices may be configured to implement the functions of the terminal, base station, or central node in the above method, and thus, the advantageous effects of the above method can also be realized. In the present application, when implementing the function of the terminal, the communication device may be one of terminals 120a to 120j shown in FIG. 1. When implementing the function of the base station, the communication device may be base station 110a or 110b shown in FIG. 1, or may be a module (for example, a chip) used in the terminal or the base station.

[0157] As shown in FIG. 9, communication device 900 includes a processing unit 910 and a transceiver unit 920. Communication device 900 is configured to implement the functions of the terminal, base station, or central node in the methods shown in FIGS. 4, 5A and 5B, 7A to 7C, or 8A to 8D.

[0158] When communication device 900 is configured to implement the function of the base station in the methods shown in FIGS. 4, 5A and 5B, 7A to 7C, or 8A to 8D, transceiver unit 920 is configured to transmit first configuration information to the terminals participating in the federated learning, where the first configuration information is used to set at least one of the following: training period, time-frequency resource, and reporting timing. The same training period, the same time-frequency resource, and the same reporting timing are set for different terminals participating in the federated learning. Also, transceiver unit 920 is configured to receive the signal obtained by wireless superposition of the gradients reported by the terminals participating in the federated learning, where the gradient is the gradient reported by the terminal at the reporting timing by using the time-frequency resource of the AI model for which the training has been completed within the training period. Processing unit 910 is configured to generate the first configuration information and process the gradients reported by the terminals.

[0159] When the communication device 900 is configured to implement the functions of a terminal in the methods shown in FIGS. 4, 5A and 5B, 7A-7C, or 8A-8D, the transceiver unit 920 is configured to transmit first configuration information from a second node, and the first configuration information is used to set at least one of the following: a training period, time-frequency resources, and reporting timing. The same training period, the same time-frequency resources, and the same reporting timing are set for different terminals participating in the federated learning. The processing unit 910 is configured to train the AI model during the training period to obtain the gradient of the AI model in the current round of model training, and the transceiver unit 920 is further configured to report the gradient of the AI model in the current round of model training to the second node at the reporting timing by using the time-frequency resources.

[0160] When the communication device 900 is configured to implement the functions of a central node in the methods shown in FIGS. 4, 5A and 5B, 7A-7C, or 8A-8D, the processing unit 910 is configured to determine second configuration information, and the second configuration information is used to set at least one of the following: a list of terminals participating in the federated learning, an initial AI model, a group temporary identifier, a training period, a terminal threshold number, the size of a transport block, or uplink requirements. The transceiver unit 920 is configured to transmit the second configuration information to the second node.

[0161] For a more detailed description of the processing unit 910 and the transceiver unit 920, refer to the relevant descriptions in the methods shown in FIGS. 4, 5A and 5B, 7A-7C, or 8A-8D. Details are not described here again.

[0162] As shown in FIG. 10, the communication device 1000 includes a processor 1010 and an interface circuit 1020. The processor 1010 and the interface circuit 1020 are coupled to each other. It can be understood that the interface circuit 1020 may be a transceiver or an input / output interface. Optionally, the communication device 1000 may further include a memory 1030 configured to store instructions executed by the processor 1010, or input data required by the processor 1010 to execute instructions, or data generated after the processor 1010 executes instructions.

[0163] When the communication device 1000 is configured to implement the above method, the processor 1010 is configured to implement the functions of the processing unit 910, and the interface circuit 1020 is configured to implement the functions of the transceiver unit 920.

[0164] When the communication device is a chip used in a terminal, the chip of the terminal implements the functions of the terminal in the above method. The chip in the terminal receives information from other modules in the terminal (for example, a radio frequency module or an antenna), and the information is transmitted to the terminal by a base station. Alternatively, the chip in the terminal transmits information to other modules in the terminal (for example, a radio frequency module or an antenna), and the information is transmitted to the base station by the terminal.

[0165] When the communication device is a module used in a base station, the module in the base station implements the functions of the base station in the above method. The module in the base station receives information from other modules in the base station (for example, a radio frequency module or an antenna), and the information is transmitted to the base station by a terminal. Alternatively, the module in the base station transmits information to other modules in the base station (for example, a radio frequency module or an antenna), and the information is transmitted to the terminal by the base station. The module in the base station here may be a baseband chip in the base station, or may be a DU or other module. The DU here may be a DU in an open radio access network (O-RAN) architecture.

[0166] When the above device is a module used in a central node, the central node modules within implements the functions of the central node in the above method. The central node The modules within are receives information from other modules in the central node (for example, a radio frequency module or an antenna), and the information is transmitted to the central node by a base station. Alternatively, the module in the central node transmits information to other modules in the central node (a radio frequency module or an antenna), and the information is transmitted to the base station by the central node. The module in the central node may be a baseband chip or other modules in the central node.

[0167] The processor of the present application may be a central processing unit (CPU), or may be any other general-purpose processor, a digital signal processor (DSP), or an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or any other programmable logic device, transistor logic device, hardware component, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0168] The memory of the present application may be a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, a register, a hard disk, a removable hard disk, a CD-ROM, or any other form of storage medium well-known in the art.

[0169] For example, the storage medium may be coupled to the processor, whereby the processor can read information from the storage medium or write information to the storage medium. The storage medium may alternatively be a component of the processor. The processor and the storage medium may be disposed in an ASIC. Further, the ASIC may be located in a base station or a terminal. Certainly, the processor and the storage medium may exist as separate components in a base station or a terminal.

[0170] Some or all of the methods of the present application can be implemented by software, hardware, firmware, or any combination thereof. When software is used to implement the method, some or all of the method may be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the procedures or functions according to the present application are completely or partially generated. The computer may be a general-purpose computer, a dedicated computer, a computer network, a network device, a user device, a core network device, an OAM, or other programmable devices. The computer program or instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer program or instructions may be transmitted in a wired or wireless manner from one website, computer, server, or data center to another website, computer, server, or data center. The computer-readable storage medium may be any useful medium accessible by a computer, or a data storage device, such as a server or data center incorporating one or more useful media. The useful medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape, or an optical medium, such as a digital video disk, or a semiconductor medium, such as a solid state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include two types of storage media, namely, a volatile storage medium and a non-volatile storage medium.

[0171] In the present application, unless stated otherwise or there is no logical contradiction, the terms and / or descriptions in different embodiments are consistent and may be cross-referenced, and the technical features in different embodiments may be combined based on their internal logical relationships to form new embodiments.

[0172] In this application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship between related objects, indicating that there may be three relationships. For example, A and / or B can represent the following cases: A exists alone, both A and B exist, and B exists alone, and A and B may be singular or plural. In the text description of this application, the character " / " usually indicates the "logical sum" relationship between related objects. In the formulas in this application, the character " / " indicates the "division" relationship between related objects. "Including at least one of A, B, or C" can mean including A, including B, including C, including A and B, including A and C, including B and C, and including A, B, and C.

[0173] It can be understood that the various numbers used in this application are only distinguished for the convenience of description and are not intended to limit the scope of this application. The serial numbers of the above processes do not mean the execution order, and the execution order of the processes should be determined based on the functions and internal logics of the processes.

[0174] [Cross-reference to related applications] This application claims the priority of Chinese Patent Application No. 202111505116.7, titled "METHOD FOR TRAINING ARTIFICIAL INTELLIGENCE AI MODEL IN WIRELESS NETWORK AND APPARATUS", filed with the China National Intellectual Property Administration on December 10, 2021, and the previous Chinese patent application is incorporated herein by reference in its entirety.

Claims

1. Transmitting first configuration information to a terminal participating in federated learning, where the first configuration information is used to set a training period, time-frequency resources, and reporting timing, and the same training period, the same time-frequency resources, or the same reporting timing is set for different terminals participating in federated learning, and Receiving a signal obtained by wireless superposition of gradients reported by the terminal participating in federated learning, where the gradient is the gradient reported by the terminal at the reporting timing by using the time-frequency resources of an artificial intelligence (AI) model for which training is completed within the training period, and A method having the above.

2. Receiving a training completion instruction from the terminal, where the training completion instruction is transmitted by the terminal to a second node when the training of the AI model is completed within the training period, and Collecting a statistical value regarding the number of terminals that have completed the training of the AI model within the training period based on the training completion instruction transmitted by the terminal, and The method according to claim 1, further having the above.

3. When the number of terminals that have completed the training of the AI model is equal to or greater than a terminal threshold number, determining an average gradient in the current round of model training based on the gradients reported by the terminals participating in federated learning, or when the number of terminals that have completed the training of the AI model is less than the terminal threshold number, using the average gradient in the previous round of model training as the average gradient in the current round of model training, and Updating parameters of the AI model based on the average gradient in the current round of model training and transmitting the average gradient in the current round of model training to the terminal, and The method according to claim 2, further having the above.

4. The method according to claim 2, further having transmitting the number of terminals that have completed the training of the AI model within the training period and the signal obtained by the wireless superposition of the gradients reported by the terminal to a first node. The method according to claim 2.

5. The first configuration information is further used to set at least one of dedicated bearer RB resources, modulation scheme, initial AI model, or transmission power. The method according to claim 1.

6. The process of determining the transmission power is Measuring a sounding reference signal (SRS) from the terminal to determine uplink channel quality of the terminal; Determining the transmission power of the terminal based on the uplink channel quality and having The method according to claim 5.

7. The first configuration information is further used to configure at least one of a dedicated bearer RB resource, a modulation scheme, an initial AI model, a channel state information (CSI) interval, or a channel inversion parameter. The method according to claim 1.

8. Further comprising receiving second configuration information from a first node, The second configuration information is used to configure at least one of a list of terminals participating in federated learning, an initial AI model, a group temporary identifier, the training period, a terminal threshold number, a transport block size, or uplink requirements. The method according to claim 1.

9. Further comprising receiving first terminal information from the terminal and transmitting second terminal information to a first node, The first terminal information includes at least one of the communication capability of the terminal, the computing capability of the terminal, or the dataset characteristics of the terminal, and the second terminal information includes at least one of the communication capability of the terminal, the computing capability of the terminal, the dataset characteristics of the terminal, or a terminal temporary identifier, and the terminal temporary identifier is assigned to the terminal by a second node. The method according to claim 1.

10. When a model training end condition is satisfied, transmitting a model training end instruction to the terminal, or Receiving a model training end instruction from a first node and transmitting the model training end instruction to the terminal The method according to claim 1, further comprising.

11. Receiving first configuration information from a second node, the first configuration information being used to configure a training period, time-frequency resources, and reporting timing, and the same training period, the same time-frequency resources, or the same reporting timing being configured for different terminals participating in federated learning; Training the AI model during the training period to obtain a gradient of the AI model in the current round of model training; Reporting the gradient of the AI model in the current round of model training to the second node at the reporting timing by using the time-frequency resources A method having.

12. When the training period ends, if the training of the AI model is completed, sending a training completion instruction to the second node The method according to claim 11, further comprising.

13. If the training of the AI model is not completed within the training period, further comprising ending the training of the AI model The method according to claim 11.

14. Receiving the average gradient in the round before model training from the second node, and updating the gradient of the AI model in the current round of model training based on the average gradient in the round before model training, or Further comprising updating the gradient and parameters of the AI model in the current round of model training based on the average gradient in the current round of model training The method according to claim 11.

15. The first setting information is further used to set at least one of dedicated bearer RB resources, modulation scheme, initial AI model, or transmission power The method according to claim 11.

16. The first setting information is further used to set at least one of dedicated bearer RB resources, modulation scheme, initial AI model, channel state information (CSI) interval, or channel inversion parameter The method according to claim 11.

17. When the first setting information is further used to set the channel state information (CSI) interval and the channel inversion parameter, the method comprises If the same frequency resource is set for the downlink channel and the uplink channel, determining the CSI of the uplink channel of the terminal based on the measured CSI of the downlink channel If the CSI of the uplink channel meets the requirements of the CSI interval, determining the transmission power based on the channel inversion parameter, and further comprising Reporting the gradient of the AI model in the current round of model training to the second node Reporting the gradient of the AI model in the current round of model training to the second node based on the determined transmission power The method according to claim 16.

18. The first setting information further includes a group temporary identifier, and the group temporary identifier is a group temporary identifier assigned to the terminal by the first node. The method according to claim 11.

19. Receiving a scheduling instruction from the second node, the scheduling instruction including a group temporary identifier, and when the group temporary identifier included in the scheduling instruction is the same as the group temporary identifier assigned to the terminal by the first node, executing the training of the AI model in the current round of the model training; or when the group temporary identifier included in the scheduling instruction is not the same as the group temporary identifier assigned to the terminal by the first node, skipping the execution of the training of the AI model in the current round of the model training The method according to claim 18, further comprising.

20. Receiving a model training end instruction from the second node, and ending the training of the AI model based on the model training end instruction The method according to claim 11, further comprising.

21. further comprising transmitting first terminal information to the second node, wherein the first terminal information includes at least one of the communication capability of the terminal, the computing capability of the terminal, or the dataset characteristics of the terminal. The method according to claim 11.

22. determining second setting information, the second setting information being used to set uplink requirements, and transmitting the second setting information to a second node to enable the second node to set a training period, time-frequency resources, and reporting timing for terminals participating in federated learning based on the uplink requirements, and receiving, from the second node, a signal obtained by wireless superposition of gradients reported by the terminals participating in federated learning, the gradients being the gradients reported by the terminals at the reporting timing by using the time-frequency resources of an artificial intelligence (AI) model whose training is completed within the training period. A method having.

23. The second setting information is further used to set at least one of a list of the terminals participating in the federated learning, an initial artificial intelligence (AI) model, a group temporary identifier, the training period, a terminal threshold number, or a size of a transport block. The method according to claim 22.

24. The second setting information is further used to set a list of the terminals participating in the federated learning, and the method includes: receiving second terminal information from the second node, the second terminal information including at least one of a communication capability of a terminal, a computing capability of the terminal, a dataset feature of the terminal, or a terminal temporary identifier, the terminal temporary identifier being assigned to the terminal by the second node; and determining the list of the terminals participating in the federated learning based on the second terminal information. The method according to claim 22, further comprising the above.

25. The second setting information is further used to set a terminal threshold number, and the method includes: receiving, from the second node, a number of terminals that have completed training of an artificial intelligence (AI) model during the training period; when the number of the terminals that have completed training of the AI model during the training period is greater than or equal to the terminal threshold number, determining an average gradient in the current round of model training based on the gradients of the AI model reported by the terminals, and when the number of the terminals that have completed training of the AI model during the training period is less than the terminal threshold number, using an average gradient in a previous round of model training as the average gradient in the current round of model training; and updating parameters of the AI model based on the average gradient in the current round of model training, and transmitting the average gradient in the current round of model training to the second node so that the second node can transmit the average gradient in the current round of model training to the terminals. The method according to claim 22, further comprising the above.

26. further comprising transmitting a scheduling instruction to the second node, the scheduling instruction including a group temporary identifier, the scheduling instruction being used to schedule terminals corresponding to the group temporary identifier to execute training of an artificial intelligence (AI) model in the current round of model training. The method according to claim 22.

27. When the model training end condition is satisfied, further comprising sending an instruction to end the model training to the second node, and instructing the terminal to end the training of the artificial intelligence (AI) model in the current round of model training. The method according to claim 22.

28. An apparatus for training an artificial intelligence (AI) model in a wireless network, An apparatus having a unit configured to implement the method according to any one of claims 1 to 10.

29. An apparatus for training an artificial intelligence (AI) model in a wireless network, Having a processor and a memory, The processor is configured to implement the method according to any one of claims 1 to 10. Apparatus.

30. An apparatus for training an artificial intelligence (AI) model in a wireless network, An apparatus having a unit configured to implement the method according to any one of claims 11 to 21.

31. An apparatus for training an artificial intelligence (AI) model in a wireless network, Having a processor and a memory, The processor is configured to implement the method according to any one of claims 11 to 21. Apparatus.

32. An apparatus for training an artificial intelligence (AI) model in a wireless network, An apparatus having a unit configured to implement the method according to any one of claims 22 to 27.

33. An apparatus for training an artificial intelligence (AI) model in a wireless network, Having a processor and a memory, The processor is configured to implement the method according to any one of claims 22 to 27. Apparatus.

34. A computer program having instructions, When the instructions are executed by a computer, the instructions cause the computer to execute the method according to any one of claims 1 to 10. Computer program.

35. A computer program having instructions, When the instructions are executed by a computer, the instructions cause the computer to execute the method according to any one of claims 11 to 21. Computer program.

36. A computer program having instructions, When the instructions are executed by a computer, the instructions cause the computer to perform the method according to any one of claims 22 to 27. A computer program. **Claim 37** A chip system having a processor or circuit configured to perform the method according to any one of claims 1 to 10. **Claim 38** A chip system having a processor or circuit configured to perform the method according to any one of claims 11 to 21. **Claim 39** A chip system having a processor or circuit configured to perform the method according to any one of claims 22 to 27.

Images